artificial-intelligence-deep-learning-neural-networks-ai.webp

Behavior trees, state machines, neural nets, steering, pathfinding, proc gen, etc, etc.

I have taken broad strokes in the field of AI/ML. Early on, it was pure ML, like image recognition and text classification. Now, I try to create novel experiences that gives users a sense of wonder and calm.

Specialized Behavior Trees

What are Behavior Trees?

A behavior tree is a data structure that controls the decision-making process of an AI. Most AI have very simple behavior trees, like the classic Idle-Chase-Attack.

Small tweaks and additions to these trees can create some very unique and specialized behaviors in enemies. Below are some examples of specialized enemy behaviors.

Note: While they may look and function similarly, behavior trees are not the same as Finite State Machine. The Player uses an FSM to control input & physics depending on their state. The bosses have an FSM on top of their decision tree. The decision tree regulates the FSM of the bosses.

The Shield

This is a very specialized enemy agent from my game. The objective of the Shield is to find someone to protect and protect them. Every other enemy agent has an assigned priority value that allows the Shield to prioritize more important enemies.

Potential Improvement: The Shield has no attack state. If the player gets too close, the Shield could bash the player, knocking them back.

The Medic

The Medic, while simpler in his behavior, dramatically changes the dynamic of gameplay. Especially when paired with the Shield (the Medic has a very high priority value).

The Medic idles until there is a dead enemy to be revived. If there are no enemies, the Medic panics and tries to attack the Player.

Potential Improvement: The Medic's idle is a stationary and very vulnerable state. I'd like to create a behavior that has the Medic find cover from the Player while they're not busy reviving enemies.

Procedural Generation

Basic Map Gen

Using BSP, I generate a map by injecting rooms into appropriate spaces and connecting them via distance from start. At first, the obstacles and decorations in the rooms were procedurally generated. Eventually we ended up using premade rooms to put into the partitioned spaces.

Extra Room Gen

Sometimes maps are generated with unused space because no rooms fit the conditions of the allocated space. So I added a final pass that fills up certain spaces. To account for these unconventional rooms, enemy spawn count is determined by the size of the room. These rooms, typically smaller (though one is extremely large in this example), spawn a specific set of enemies.

BSP + Cellular Automata

Here is an example of an early iteration of our map generation. I was trying to settle on a direction for the game, so I created several proof of concepts. The idea: an abandoned prison overgrown by the surrounding forest.

Steering & Pathfinding

Flock Steering

Here I've created a proof-of-concept regarding steering. To the left I've called the Pariah Scenario. The Pariahs are the independent reds that act as influencing agents over the rest of the flock.